[mpi3-coll] array parameters in nonblocking collectives

[Torsten Hoefler]

Hello Bin,
thanks for bringing this up Bin, it seems like an important issue.

I can see both sides. In general, I would say that all vector
collectives are not scalable anyway because they require \Theta(P)
memory per node and \Theta(P^2) total. Shifting this by a constant
factor probably doesn't matter too much for asymptotic (large-scale)
analysis. However, there is certainly a border-zone where something
could be gained with transferring the ownership of the arrays. 

We will certainly discuss this issue in Helsinki and at the next telecon!

For the future, I would generally expect that, at large scale, such
communications are rather sparse. So we should either talk about a
sparse interface to vector collectives or its static variant, the graph
topology collectives. Those options will also be discussed asap. Both
options would potentially reduce the total memory consumption (assuming
sparse communications) from \Theta(P^2) to O(P) or O(P log(P)) depending
on the problem.

All the Best,
  Torsten

-- 
 bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Torsten Hoefler       | Postdoctoral Fellow
Open Systems Lab      | Indiana University    
150 S. Woodlawn Ave.  | Bloomington, IN, 474045, USA
Lindley Hall Room 135 | +01 (812) 855-3608